Evaluation of advanced automatic PET segmentation methods using nonspherical thin-wall inserts.

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Image Segmentation

Background:

• Positron emission tomography (PET) segmentation is vital for radiotherapy planning.
• Current validation of PET automatic segmentation (PET-AS) methods often uses simplified spherical phantoms.
• The effect of complex, nonspherical geometries on PET-AS accuracy remains underexplored.

Purpose of the Study:

• To evaluate and compare eight promising PET-AS methods.
• To assess the impact of nonspherical, thin-wall insert geometries on segmentation accuracy.
• To generate clinically realistic data for robust PET-AS validation.

Main Methods:

• Manufactured 16 nonspherical thin-wall inserts (ellipsoids, toroids, tubes, pears, drops) and scanned them in a phantom.
• Used a baseline study with six spherical inserts of varying volumes.
• Applied eight in-house PET-AS methods (adaptive iterative thresholding, region-growing, clustering, gradient-based) to acquired images.
• Measured delineation accuracy using Dice Similarity Coefficient (DSC) against CT contours and dimensional errors.

Main Results:

• Delineation accuracy was lower for nonspherical inserts compared to spheres in 88% of cases.
• Gradient-based methods showed poor performance on tori (DSC < 0.5).
• Region-growing achieved high accuracy (DSC > 0.76) but had significant dimensional errors for pear/drop shapes.
• Adaptive iterative thresholding demonstrated the highest DSC ( > 0.83) on nonspherical data, showing geometric robustness.

Conclusions:

• The study confirms that nonspherical geometries significantly affect PET-AS accuracy, revealing weaknesses not seen with spherical phantoms.
• Adaptive iterative thresholding proved robust, while other methods like gradient-based and region-growing showed specific limitations.
• Utilizing diverse, complex geometries is essential for thorough validation and further development of PET-AS tools.